Models & Labs
Underdog demos 32MB local speech model
TechCrunch AIhigh confidence
Why it matters
- →Proves high-fidelity voice generation is possible with extremely small model sizes (32MB).
- →Enables zero-latency, offline audio processing directly in web browsers.
- →Offers a privacy-first alternative to cloud-based speech APIs for local applications.
Underdog has released 'Woof,' a 32MB speech generation model capable of running locally in web browsers without internet connectivity. The startup demonstrated the model's capabilities through an interactive browser game where users can type text to be spoken aloud or generate music via a virtual tuner, all processed on-device. This follows their previous release of Bark, a 167MB speech-to-text model, as part of a broader strategy to provide open, private AI alternatives to cloud-dependent services. Founder Sigil Wen aims to enable users to own their data by running capable models on existing hardware.
Read originalThe story around this
Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
How we got here
- 1Fish Audio raises $52M for AI voice models
TechCrunch AI · July 28, 2026 · Background
- 2Smallest.ai raises $13M for human-like voice AI
TechCrunch AI · July 31, 2026 · Background
- 3BreezeTTS2: Real-time local voice synthesis
Sam Witteveen · August 31, 2026 · Related
- 4Block launches Dessa, a speech-to-speech AI model
The Verge AI · October 23, 2026 · Background
- 5llama.cpp adds Soprano TTS support
llama.cpp Releases · October 24, 2026 · Background
- 6llama.cpp adds LiquidAI LFM3-Audio and Omni support
llama.cpp Releases · October 28, 2026 · Background
- 7Hugging Face brings open ASR and decision models to Swift
Hugging Face Blog · October 29, 2026 · Background
- 8Phonon TTS: High-Quality On-Device Speech
Hugging Face Blog · November 5, 2026 · Background
More from TechCrunch AI
© TechCrunch AIMarket & Regulationbusiness
Google buys Crusoe's Wyoming data center project
Google is acquiring an unfinished data center from Crusoe in Wyoming, signaling a strategic pivot to direct infrastructure control amid scaling pressures. The $17 billion investment highlights the immense capital required for AI compute, as independent developers struggle with cost and timeline overruns. This move illustrates the growing vertical integration of tech giants into physical infrastructure, reducing reliance on third-party providers like Crusoe for massive deployments. By taking ownership, Google bypasses the bottlenecks that often plague external development teams. The deal reflects a broader industry trend where hyperscalers reclaim control over their most critical assets. Independent builders now face stiffer competition from companies with deeper pockets and direct operational oversight.
TechCrunch AI
© TechCrunch AIInvestment · $11B
Market & Regulationbusiness
OpenAI raises $11B at $1.25T valuation
OpenAI is securing an additional $11 billion in a funding round that values the company at $1.25 trillion, falling short of its initial $30 billion target but still signaling massive investor confidence. This latest capital injection comes as the broader AI market faces headwinds, with Anthropic’s potential IPO valuation also cooling from earlier highs. The move highlights OpenAI's continued ability to attract deep-pocketed support despite regulatory uncertainties and political shifts affecting data center expansion. While the round hasn't fully closed, the sheer scale of this valuation reinforces OpenAI's dominant position in the race for AI supremacy. Investors are betting big on the lab's ability to maintain its lead in compute and model capabilities. The delay in IPO plans suggests a strategic pause to navigate these complex market conditions. This funding ensures OpenAI remains the financial heavyweight in the sector.
TechCrunch AI
© TechCrunch AIMarket & Regulationother
Hundreds of Artex AI hacking tools exposed
Censys uncovered over 800 publicly accessible instances of Artex, an autonomous penetration testing tool linked to major breaches in Korean finance and ConnectWise. These exposed deployments reveal a critical security gap: powerful AI agents running without authentication turn offensive capabilities into open targets for misuse. The developers recently shifted to closed-source distribution, yet the code remains widely available on public repositories, leaving the threat landscape largely unchanged. This situation demonstrates how easily sophisticated AI-driven attacks can be launched by anyone with basic configuration skills. The widespread exposure underscores significant risks associated with misconfigured AI security tools.
TechCrunch AI
More in Models & Labs
Models & Labsother
llama.cpp b12177 adds CUDA 13 and ROCm 10
This release quietly closes the hardware gap for local inference by adding default support for CUDA 13.5 and ROCm 10.0 alongside existing CUDA 12 builds. AMD GPU users finally get first-class parity with NVIDIA, while Snapdragon mobile chips gain dedicated NPU acceleration paths. Apple Silicon KleidiAI is explicitly disabled here, a notable regression for Mac performance that suggests ongoing tuning issues. The inclusion of OpenVINO and SYCL backends further cements llama.cpp as the universal runtime for inference across nearly every major silicon architecture.
llama.cpp Releases
Models & Labsmodels
llama.cpp adds Kimi K3.1-Pro support
Kimi’s new K3.1-Pro model joins the llama.cpp ecosystem, bringing a distinct architectural twist to local inference. By replacing full-attention MLA layers with DSA (Dynamic Sparse Attention) and linear-attention KDA layers, this release tests how well sparse caching handles hybrid attention patterns. The implementation keeps indexer precision high to avoid quality loss during top-k selection, ensuring the model runs accurately on consumer hardware. This expands the library’s reach into newer Chinese open-weight models that rely on complex memory management.
llama.cpp Releases
Models & Labsother
llama.cpp b12182 release with CUDA 13 and ROCm 10
This release quietly cements llama.cpp as the universal inference runtime by adding default builds for CUDA 13.5 and ROCm 10.0 across Linux and Windows. The inclusion of ROCm 10 means AMD GPU users no longer need custom compilation to access modern acceleration, effectively closing the gap with NVIDIA's ecosystem support. While KleidiAI on Apple Silicon is disabled in this build, the broader platform coverage ensures that local inference remains accessible on virtually any hardware configuration without manual setup.
llama.cpp Releases